This article sets out a taxonomy of empirically observed failure modes in AI-generated corporate narratives, derived from controlled, repeatable testing across multiple large language models. The taxonomy does not rely on anecdotal incidents, post-hoc reconstruction, or hypothetical scenarios. It is organized around evidentiary consequences under scrutiny, not technical error classification.